Post by Turing

Benchmark scores are climbing at an incredible pace. But are AI models actually getting better at helping scientists do scientific work? At [ICML] Int'l Conference on Machine Learning 2026 in Seoul, Turing's Charlotte Tao will explore one of the biggest challenges in frontier AI: the growing gap between benchmark performance and real-world scientific capability. In just one year, SciCode scores increased from 4.6% to 59%, while HLE rose from 8% to 47%. Those are impressive gains, but strong benchmark results do not always translate into success across complete scientific workflows. Drawing on Turing's work developing frontier data, Charlotte will share why evaluating AI requires looking beyond isolated, auto-gradable tasks and toward the complex, end-to-end workflows that drive scientific discovery. The session will also look ahead at what's next for frontier AI evaluation, including how the field may evolve beyond static datasets toward modular, composable capability infrastructure. If we're building AI to accelerate science, we need to measure what matters. Proud to see Turing contributing to this important conversation at ICML 2026. 𝐀𝐝𝐯𝐚𝐧𝐜𝐢𝐧𝐠 𝐅𝐫𝐨𝐧𝐭𝐢𝐞𝐫 𝐒𝐜𝐢𝐞𝐧𝐭𝐢𝐟𝐢𝐜 𝐂𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬, 𝐓𝐨𝐝𝐚𝐲 𝐚𝐧𝐝 𝐓𝐨𝐦𝐨𝐫𝐫𝐨𝐰 𝐂𝐡𝐚𝐫𝐥𝐨𝐭𝐭𝐞 𝐓𝐚𝐨, 𝐓𝐮𝐫𝐢𝐧𝐠 𝐈𝐂𝐌𝐋 2026 | 𝐒𝐞𝐨𝐮𝐥 | 𝐉𝐮𝐥𝐲 6-11